Best practices for software effort estimation can include the use of
automatic techniques to summarize past data. There exists a large and
growing number of techniques. Which are useful? In this study, 158
techniques were applied to some COCOMO data. 154 158 = 97% of the
variants explored below add little or nothing to a standard linear
model (with simple column and row pruning). For example, learners that
assume multiple linear models (such as model trees) or no parametric
form at all (such as nearest neighbor estimation) perform relatively
poorly.  Also, elaborate searches did not add value to effort
estimation. Exponential time searching for the best subsets of the
columns performed no better than near-linear-time column pruning. It
is possible that other techniques, not included in the 158 techniques
studied here, might perform better. However, based on current
evidence, this study concludes that when technique learners are used
for effort estimation, a linear model with simple column and row
pruning will suf?ce (at least, for COCOMO-style data).